Predicting Small Molecule Transfer Free Energies by Combining Molecular Dynamics Simulations and Deep Learning

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Accurate prediction of small molecule hydrophobicity is essential for drug discovery, guiding processes like membrane crossing and drug-protein binding.
• Atomistic molecular dynamics (MD) simulations offer accurate free energy calculations but are computationally intensive.
• Existing machine learning (ML) and empirical methods often rely on experimental data, limiting their applicability.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting small molecule hydrophobicity using atomistic MD simulation data.
• To create a large dataset of 15,000 small molecule free energies of transfer from water to cyclohexane.
• To assess the accuracy and transferability of different ML architectures for cheminformatics modeling.

Main Methods:

• Performed atomistic molecular dynamics (MD) simulations to calculate free energies of transfer for 15,000 small molecules.
• Trained various machine learning models, including spatial graph neural networks, 3D-convolutional neural networks, and shallow learning models based on chemical fingerprints.
• Evaluated model performance using mean absolute error and tested transferability across diverse molecular sets.

Main Results:

• A spatial graph neural network achieved the highest accuracy in predicting free energies of transfer, with a mean absolute error of approximately 4 kJ/mol.
• 3D-convolutional neural networks also showed high accuracy, while shallow learning models were less effective.
• Including MD simulation data and employing multitask learning significantly improved prediction accuracy and model transferability.

Conclusions:

• Machine learning models, particularly spatial graph neural networks, can accurately predict small molecule hydrophobicity using data from atomistic MD simulations.
• The developed dataset and models offer valuable insights into hydrophobicity prediction and advance ML cheminformatics.
• This work provides a robust dataset for developing and testing future ML methods in drug discovery and chemical informatics.